acceptodds
Under review as a conference paper at ICLR 2027

AI-generated number sequences contain detectable, non-human biases

Abstract

State-of-the-art AI detection models are classifiers trained on millions of human and AI documents. To tell the difference between human and AI text, trained experts rely primarily on features like characteristic vocabulary, sentence structure, or other visible heuristics. Whether trained classifiers actually depend on these cues, however, is unclear. We test if AI detectors can operate in a domain where none of these heuristics apply: number sequences. Pangram, a state-of-the-art AI detector, distinguishes LLM-generated sequences from identically formatted human-written and true-random baselines with an AUROC of 0.995. Transformation effects are heterogeneous: reversal and some structured recodings preserve detectability, while shuffling and some bijective substitutions eliminate it, implicating an interaction between sequence arrangement and representation. Examining sequence structure across eight open models, we find that LLM-generated sequences differ from human-generated sequences in their statistical structure, including average jump size and complete 10-digit windows. A fitted, interpretable simulation model captures 84% of this structure, and the simulated sequences begin to reproduce AI scores. Sweeping checkpoints, decoding, and chat formatting, we find that post-training, templates, and truncated decoding generally amplify this structure. Thus, Pangram can exploit subtle statistical fingerprints in number sequences when lexical and syntactic cues are absent.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI Detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jHeKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work tests whether an AI-text detector can identify LLM output with simple numerical outputs. Using Pangram, FastDetectGPT and other methods, they identify this domain to contain large distributional differences from how humans traditionally choose numbers.

Critical strengths
  • Number sequences is a pretty interesting domain to consider; the number of experiments and breadth of experiments are very much appreciated.
Critical weaknesses
  • The paper is pretty descriptive. It discusses what changes causes the detector's score to change - very much like a brute force adversarial search. Most experiments are based on next-digit fit rather than on which mechanism drives detection. I think this is a pretty good empirical study but it does not help the reader understand why exactly these distributional difference arises in LLMs.
Questions for authors
  • Do the authors have an intuition on why do spaces or newlines remove about 90% of the signal?
  • I'd also love to see which phase of training / data causes this distributional difference in number sequences.
2 comments

Hello there, thank you for taking the time to consider our paper. We would like to sincerely apologize for the late response within the discussion period.

Q1: Intuition on New Lines and Spaces

We thank the reviewer for raising this question. While we observe that spaces and newlines substantially reduce the detection signal, we do not currently have a clear mechanistic explanation for this effect.

Q2: OLMo Phase Experiment

Thank you for your suggestions! We conducted an experiment based on numbers generated by different checkpoints of publicly released OLMo 2, covering intermediate pretraining checkpoints where available, final base models, SFT and DPO checkpoints, and final instruction-tuned models.

We held sampling fixed at temperature 0.7 and top-p 0.8, evaluated continuation and chat prompts separately, and measured eight sequence statistics alongside Pangram 3.3.2 scores. We collected 100 complete 100-digit sequences per successful condition, totaling 1,000 sequences.

The training stages are as follows:

  • Pretraining (1.0T tokens): Random initialization followed by pretraining on OLMo-Mix-1124. Snapshot at 999B tokens.
  • Pretraining (3.9T tokens): Same pretraining run as the 1.0T checkpoint, continued on OLMo-Mix-1124. Snapshot at 3.896T cumulative tokens.
  • Final Base: End of stage-1 pretraining followed by three stage-2 runs, each trained on 50B tokens from Dolmino-Mix-1124. Their weights were merged to form Final Base.
  • SFT: Final Base followed by supervised fine-tuning on instruction–response demonstrations from the OLMo-specific Tülu 3 mixture.
  • DPO: SFT followed by direct preference optimization using pairs of preferred and rejected responses.
  • Final Instruct: DPO followed by reinforcement learning with verifiable rewards, using PPO to update the model.

The results are as below:

Table 1: Detectability and digram entropy

\textbf{Training stage} & \textbf{Pangram mean} & \textbf{Digram entropy (nats)} \\\\ \hline \text{Pretraining (1.0T)} & 0.041 / 0.094 & 1.65 / 1.66 \\\\ \text{Pretraining (3.9T)} & 0.195 / 0.223 & 1.56 / 1.53 \\\\ \text{Final Base} & 0.168 / 0.038 & 1.57 / 1.68 \\\\ \text{SFT} & 0.908 / 0.889 & 0.43 / 0.97 \\\\ \text{DPO} & 0.867 / \text{---} & 0.17 / \text{---} \\\\ \text{Final Instruct} & 0.923 / \text{---} & 0.12 / \text{---} \end{array}$$ **Table 2: Step behavior and autocorrelation** $$\begin{array}{l|c|c} \textbf{Training stage} & \textbf{Mean step} & \textbf{Lag-1 correlation} \\\\ \hline \text{Pretraining (1.0T)} & 3.09 / 3.30 & -0.101 / -0.090 \\\\ \text{Pretraining (3.9T)} & 3.48 / 3.55 & -0.127 / -0.116 \\\\ \text{Final Base} & 2.99 / 2.74 & -0.079 / -0.046 \\\\ \text{SFT} & 3.77 / 3.71 & -0.168 / -0.142 \\\\ \text{DPO} & 3.88 / \text{---} & -0.183 / \text{---} \\\\ \text{Final Instruct} & 3.92 / \text{---} & -0.196 / \text{---} \end{array}$$ **Table 3: Coverage and digit preference** $$\begin{array}{l|c|c} \textbf{Training stage} & \textbf{Coverage (draws)} & \textbf{Favorite digit (\\%)} \\\\ \hline \text{Pretraining (1.0T)} & 83.82 / 37.37 & 18.11 / 15.41 \\\\ \text{Pretraining (3.9T)} & 10.38 / 19.72 & 16.16 / 19.16 \\\\ \text{Final Base} & 42.66 / 73.74 & 18.38 / 19.53 \\\\ \text{SFT} & 10.00 / 10.00 & 12.02 / 11.99 \\\\ \text{DPO} & 10.00 / \text{---} & 10.59 / \text{---} \\\\ \text{Final Instruct} & 10.00 / \text{---} & 10.89 / \text{---} \end{array}$$ **Table 4: Repetition and step behavior** $$\begin{array}{l|c|c} \text{Training stage} & \pm 1\\ \text{steps }(\\%) & \text{Immediate repeats }(\\%) \\\\ \hline \text{Pretraining (1.0T)} & 21.03 / 21.49 & 4.33 / 1.08 \\\\ \text{Pretraining (3.9T)} & 19.45 / 18.93 & 2.81 / 4.60 \\\\ \text{Final Base} & 21.68 / 21.08 & 6.90 / 11.88 \\\\ \text{SFT} & 14.62 / 18.95 & 0.28 / 0.01 \\\\ \text{DPO} & 11.32 / \text{---} & 0.19 / \text{---} \\\\ \text{Final Instruct} & 11.25 / \text{---} & 0.22 / \text{---} \end{array}$$ **Table 5: Complete ten-digit block structure** $$\begin{array}{l|c} \textbf{Training stage} & \textbf{Complete 10-digit blocks (\\%)} \\\\ \hline \text{Pretraining (1.0T)} & 1.10 / 7.10 \\\\ \text{Pretraining (3.9T)} & 9.20 / 7.80 \\\\ \text{Final Base} & 5.40 / 0.30 \\\\ \text{SFT} & 61.70 / 53.00 \\\\ \text{DPO} & 87.70 / \text{---} \\\\ \text{Final Instruct} & 87.50 / \text{---} \end{array}$$ *Each cell reports continuation / chat. “—” indicates that no complete 100-digit sequence was obtained within 500 generation attempts.* From this OLMo 2 experiment, we can see that distributional biases are already present during pretraining and become substantially stronger during the Base-to-SFT transition, alongside a sharp increase in detectability. These biases persist through DPO and RLVR, suggesting that SFT is a major stage of amplification.
Reviewer_jHeKOpenReview ↗

Response to Authors

Thank you for the detailed rebuttal. I think something worth adding to this paper would be some sort of mechanistic study on these biases and why exactly distribution differences arose, as stated in my weaknesses.

As a result, I will keep my score.

Reviewer_fPP7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether AI-generated random number sequences contain statistical patterns that allow AI-text detectors to distinguish them from human-generated and truly random sequences. The authors show that existing detectors can identify LLM-generated number sequences with high accuracy, even when lexical and syntactic information is absent.

The paper also studies how detectability changes under transformations of the sequences and different generation settings. Finally, the authors propose an interpretable simulation model that captures 84% of the non-random structure in single-digit sequences and partially reproduces their AI detection scores.

Overall, I find the paper interesting and the main observations valuable. While the experimental analysis has some limitations, I believe the work raises an important question and provides useful evidence toward understanding the statistical signatures of AI-generated content. I lean toward a borderline accept and would welcome further analysis of the three-digit sequences and frontier models.

Critical strengths

I like this paper. It studies an unusual and interesting problem that could lead to important insights into AI-generated text detection.

The main result is surprising: even sequences of random numbers contain enough information to distinguish AI-generated content from human-generated or truly random content.

I particularly appreciate the experiments with sequence transformations. They show that detectability depends not only on the values appearing in a sequence, but also on their order and representation.

Another strong part of the paper is the interpretable simulation model. The authors identify several statistical biases in LLM-generated sequences and show that a relatively simple model can reproduce much of this structure. This provides a useful step toward understanding what makes AI-generated content detectable.

Critical weaknesses

My main concern is that the most interesting detection results are obtained for three-digit numbers, while the detailed statistical analysis and simulation model are developed only for single-digit sequences. The three-digit setting shows considerably higher detectability, and I would like to understand which statistical properties explain this difference.

Another limitation is the distribution of models in the experiments. Most of the detailed analysis is based on relatively small and older open-weight models, while the four frontier models contribute only a small number of sequences. The aggregated results therefore do not provide a sufficiently clear picture of how the observed patterns differ between small and frontier LLMs.

Finally, most experiments rely on a single commercial detector, Pangram, whose implementation is not publicly available. Although the authors also evaluate Fast-DetectGPT and Binoculars and report strong detection performance, the analysis of sequence transformations and the simulation model is almost entirely based on Pangram. It is therefore unclear to what extent the conclusions generalize to other detectors.

Questions for authors

Could the authors extend the statistical analysis and simulation model to three-digit number sequences? This seems particularly important because these sequences have higher detection scores. I would be interested to see which statistical patterns explain their detectability and whether the same mechanisms found for single-digit sequences are sufficient.

Could the authors report more results separately for small open-weight models and frontier LLMs? In particular, it would be useful to understand whether the same biases are present in frontier models and how their strength changes with model size and generation settings.

Could the authors extend at least some of the transformation experiments to Fast-DetectGPT and Binoculars? This would help establish whether the observed effects reflect general properties of AI-generated sequences or are specific to Pangram.

1 comment

Thank you for your review and for carefully considering our work! We sincerely apologize for the late response during the discussion period. Nonetheless, we appreciate the opportunity to address your questions and concerns!

Q1: Statistical Modeling of Three-Digit Sequences

We thank the reviewer for their suggestion here.

We focus our mechanistic analysis on single digits because the constrained output space is what makes it possible. Under uniform sampling, a 100-item list contains about ten occurrences of each digit, and there are only 100 possible types of transitions between consecutive digits. Therefore, digit preference, repetition avoidance, and window-level coverage can be estimated per list.

With 1,000 values, each value would only appear 0.1 times per list on average (in the random setting), with 10610^6 possible transition types, which makes the deeper statistical analysis and simulation we conducted more challenging. We include three-digit sequences in the detection and transformation experiments because we believe the relationship between cardinality and detectability is of interest.

Q2: Do frontier LLMs share the same biases?

Here, we expand Table 3 to include the mix of Frontier LLMs used throughout the rest of the work.

\textbf{Statistic (chance value)} & \textbf{True RNG} & \textbf{Human} & \textbf{Open LLMs} & \textbf{Frontier LLMs} \\\\ \hline \text{Immediate-repeat rate (10\\%)} & 10.1\\% & 10.6\\% & 0.0\\% & 0.8\\% \\\\ \text{10-digit windows using every digit (0.04\\%)} & 0.1\\% & 0.4\\% & 62.6\\% & 90.7\\% \\\\ \text{Steps until all ten digits seen (28)} & 27.7 & 62.0 & 10.7 & 11.2 \\\\ \text{Most-used digit's share (15\\%)} & 15.1\\% & 19.3\\% & 11.5\\% & 10.4\\% \\\\ \text{Mean step size (3.3)} & 3.27 & 2.16 & 3.66 & 4.13 \\\\ \text{Lag-1 autocorrelation (0)} & -0.01 & +0.34 & -0.13 & -0.30 \\\\ \text{±1 step rate (18\\%)} & 18.0\\% & 40.7\\% & 19.9\\% & 8.4\\% \end{array}$$ We find that frontier models indeed show the same deviations from randomness as the open models! ### Q3: Fast-DetectGPT and Binoculars Across Transformations Thank you for your suggestion! We extended our evaluation of Fast-DetectGPT and Binoculars on the exact same three-digit sequences and transformed texts as in Table 13, covering all 44 transformations. The results show that some transformation effects are shared across detectors, while others are more detector-dependent. Following the retained/eliminated-signal measure in Table 1, we report the **LLM–RNG mean score gap after transformation as a percentage of the original gap**. **Shared transformation effects:** $$\begin{array}{l|l|r|r|r} \textbf{Transformation} & \textbf{Shared effect} & \textbf{Pangram} & \textbf{Fast-DetectGPT} & \textbf{Binoculars} \\\\ \hline \text{Atbash} & \text{Largely retained} & 101.1\\% & 99.5\\% & 77.8\\% \\\\ \text{Remove leading zeros} & \text{Largely retained} & 74.5\\% & 85.5\\% & 70.7\\% \\\\ \hline \text{Simple substitution} & \text{Nearly eliminated} & 0.3\\% & 7.1\\% & 9.1\\% \\\\ \text{Autokey cipher} & \text{Nearly eliminated} & 1.4\\% & 2.3\\% & -9.7\\% \\\\ \text{Enigma-style rotor cipher} & \text{Nearly eliminated} & 0.0\\% & 12.4\\% & 6.8\\% \\\\ \text{Five-rail rail fence cipher} & \text{Nearly eliminated} & 1.7\\% & 12.1\\% & -2.0\\% \\\\ \text{Playfair cipher} & \text{Nearly eliminated} & 0.3\\% & 10.6\\% & -11.4\\% \end{array}$$ **Detector-dependent transformation effects:** $$\begin{array}{l|r|r|r} \textbf{Transformation} & \textbf{Pangram} & \textbf{Fast-DetectGPT} & \textbf{Binoculars} \\\\ \hline \text{Reverse the list} & 100.2\\% & 41.6\\% & 41.5\\% \\\\ \text{ROT5} & 101.5\\% & 15.7\\% & -7.5\\% \\\\ \text{Full-width digits} & 98.8\\% & 16.4\\% & -46.2\\% \\\\ \text{Substitute hundreds digit only} & 85.8\\% & -6.1\\% & -19.9\\% \\\\ \text{JSON formatting} & 80.4\\% & -5.5\\% & 4.8\\% \\\\ \text{Square each value} & 3.6\\% & 144.9\\% & 232.9\\% \\\\ \text{Sort the list} & 0.4\\% & 122.9\\% & 208.5\\% \\\\ \text{Base64 encoding} & 0.5\\% & 100.8\\% & 102.7\\% \\\\ \text{Beaufort cipher} & -0.4\\% & 93.8\\% & 143.6\\% \end{array}$$ Fast-DetectGPT and Binoculars exhibit generally similar responses across transformations, while their agreement with Pangram is substantially lower. **Correlation across transformations:** $$\begin{array}{l|r|r} \textbf{Detector pair} & \textbf{Pearson }r & \textbf{Spearman }\rho \\\\ \hline \text{Fast-DetectGPT--Binoculars} & 0.914 & 0.923 \\\\ \text{Pangram--Fast-DetectGPT} & 0.154 & 0.129 \\\\ \text{Pangram--Binoculars} & 0.044 & 0.026 \end{array}$$ This agreement is consistent with their shared reliance on language-model probabilities. Fast-DetectGPT measures conditional probability curvature, while Binoculars uses perplexity-based scoring. In contrast, Pangram uses a Transformer classifier trained through supervised fine-tuning for detection.
Reviewer_HhWsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether general-purpose AI text detectors can identify number sequences generated by language models when common clues from vocabulary, sentence structure, and meaning are removed. It also examines which statistical patterns are linked to detectability.

The authors collect sequences from humans, language models, a pseudorandom number generator, and a true random source. Pangram achieves an AUROC of 0.995 in the specified single-digit experiment. Two other detectors also show strong performance. Experiments with different output sets and sequence transformations suggest that detectability depends on both the order of the items and how they are represented. Controlled experiments with open-weight models further show that instruction tuning, chat templates, and the default sampling settings used in the paper generally strengthen sequence patterns and increase detectability.

Further analysis shows that some models strongly avoid repeating recently used digits. They also frequently produce blocks of ten digits that contain each digit from 0 to 9 exactly once. Based on these findings, the authors fit a separate, interpretable simulation with 66 parameters for each of eight models. On average, these simulations capture 84% of the non-random structure under the paper’s KL-based measure. Their outputs receive an average Pangram score of 0.63, compared with 0.80 for the original model outputs.

The main contribution is to identify detectable statistical patterns in number sequences and connect some of these patterns to detection scores through behavioral simulations.

Critical strengths
  1. Number sequences provide a simple, controlled setting for this research. Matching sequence length and format helps reduce common language clues. This allows the authors to focus on statistical patterns linked to detectability. This research approach offers some novelty.

  2. The simulations include digit preferences, avoidance of recently used digits, position-based patterns, and pattern copying. They are fitted without using Pangram scores as a training target. Their outputs still receive relatively high detection scores. This provides evidence of a link between these behaviors and detectability.

Critical weaknesses
  1. The practical value of the findings remains unclear. The experiments mainly focus on random sequences of a fixed length. The paper does not yet show how these findings could help with real tasks, such as improving AI text detection, explaining detection errors, or assessing a model’s ability to generate random outputs. This makes the practical impact difficult to assess.

  2. The explanation of the detector’s decisions remains incomplete. The paper identifies several statistical biases and reproduces part of the detection signal through simulations. However, it does not clearly establish how much each mechanism contributes to detection scores. It is therefore difficult to determine how well the proposed patterns explain the detector’s decisions, rather than simply being associated with high scores.

Questions for authors
  1. What specific practical applications could this research support? How do the current experimental results support these applications?

  2. Could the generation and filtering process introduce selection bias? Please report the number of generation attempts, the proportion of valid sequences, and the exclusion rules for each setting. Please also explain whether these differences affect the main conclusions.

  3. Which simulation mechanisms have the greatest effect on detection scores? Could you remove each mechanism in turn to measure its contribution? In particular, please assess how avoidance of recent digits, ten-digit block structure, and pattern copying affect both the fit to the model’s probability distribution and the detection scores.

1 comment

Hello there! The authors would like to sincerely apologize for the late submission during the discussion period. We appreciate your careful reading of our work and are glad to address the following questions.

Q1: On Practical Applications

We believe that the practical applications of this research lie in the field of AI detector explainability. An emerging direction in AI detection research investigates the features that detectors use to make their decisions. This work often relies on automated interpretability pipelines or manual labeling to explain detections through visible cues, such as "AI vocabulary" or formulaic structure. Examples include:

  • Kuznetsov et al. (2025), Feature-level insights into artificial text detection with sparse autoencoders.
  • Creo & Ranganath (2026), Show, Don't TELL: Explainable AI-generated text detection (arXiv:2605.27921).
  • Pudasaini et al. (2026), Why AI-generated text detection fails: Evidence from explainable AI beyond benchmark accuracy.

Our results demonstrate that detectors can still distinguish AI-generated content when these visible cues are absent. This invites a shift in methodology toward investigating less obvious statistical and sequential features that may contribute to detector decisions, rather than relying exclusively on human-interpretable stylistic markers.

Q2: On Generation and Filtering Mechanisms

In the paper, our procedure was to generate open-model sequences in batches of 20, continuing until we reached our target number of accepted sequences per setting. A generated list was accepted if it contained at least 100 parsed digits. If a sequence contained more than 100 digits, we retained the first 100 parsed digits and discarded the remainder. Outputs that failed any check were excluded. The OLMo Instruct + Chat settings were omitted from the final report because they never yielded an eligible list across 39 batches.

Our original dataset generation runs used repeated batch attempts until the target was reached. Because the overall batch count and validity rate were not recorded, we cannot report an attempt count or acceptance proportion for every setting in our dataset.

However, we conducted an additional pilot filtering-validity experiment using Qwen 2.5 0.5B. We generated 200 attempts per setting and recorded the acceptance rates.

SettingPure samplingDefault sampling
Base, continuation105/200 (52.5%)200/200 (100%)
Base, chat40/200 (20%)198/200 (99%)
Instruct, continuation105/200 (52.5%)200/200 (100%)
Instruct, chat5/200 (2.5%)187/200 (93.5%)

Across settings, acceptance rates ranged from 2.5% to 100%. Settings with low acceptance rates necessarily yield a more selected subset of model outputs, and we recognize the potential selection effects this introduces. Our results are intended to characterize patterns among valid, 100-digit sequences, which certain models and settings are more likely to produce. We will add a description of our filtering procedure to the final version, along with an acknowledgment of differences in validity rates and their potential implications.

Q3: Simulator Ablations

Thank you for your suggestion! We are glad to present an additional experiment addressing the reviewer's interest in understanding which mechanisms contribute to the kernel's fit and its AI detection score. The results are presented below.

GeneratorStructure capturedAI score
Simulation model84%0.63
Simulation model refit with one mechanism removed:
− Pattern copying76%0.23
− Recent-digit avoidance80%0.53
− Ten-digit block structure81%0.63
Real LLM lists100%0.80

Pattern copying appears to be the primary driver of AI detection scores among the mechanisms considered. Removing this mechanism lowers the mean AI score from 0.63 to 0.23, while also reducing the structure captured from 84% to 76%.

Removing recent-digit avoidance decreases both the structure captured (by 4 percentage points) and the mean AI score (by 0.10). In contrast, removing the ten-digit block structure decreases the structure captured by 3 percentage points while leaving the mean AI score unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.